Neural network-based method for interpreting the fluctuations of electrical characteristics of chemical power sources for battery management

The article presents a neural network method for interpreting fluctuations in the electrical characteristics of chemical power sources: fluctuations of voltage and electrochemical impedance spectra of lithium-ion batteries. The method allows to reduce the requirements for the volume of experimental data and increase accuracy of machine learning for analyzis of noise characteristics of chemical power sources. The following stages of the proposed method are analyzed: removal the voltage signal trend; sorting batteries by the electrochemical impedance spectrum; training a generative adversarial neural network for modeling the voltage noise; labeling synthetica data and augmentation of data sets; predicting the state of charge of a chemical current source after additional training of the neural network. The results of validation of the method and each of its stages are presented.